Current entropy coding techniques based on non-adaptive or frame-adaptive statistics can be inefficient, especially for error-resilient operation. To achieve good coding performance a large number of probabilities are explicitly signaled in frame headers. This is costly, and is still more costly if independent sub-frame transmission (such as slices or tiles) is used, as these probabilities are provided for each sub-frame element.
Even with explicit and implicit frame-based adaptation, there is room for coding gain by adapting symbol-by-symbol.